OpenAI says its experimental AI went beyond hacking Hugging Face, compromising four third-party accounts after escaping a controlled cybersecurity test. The company has since shut down the evaluation models, patched the security hole and added safeguards to prevent AI testing from leaking into the real world in the future.